Summary
- The FSB has opened consultation on 12 sound practices for responsible AI adoption in financial institutions.
- The proposed practices are intended to support board and senior management oversight across the AI lifecycle.
- The consultation covers newer forms of AI, including generative and agentic AI, alongside cyber resilience and third-party risk.
The Financial Stability Board has opened consultation on proposed sound practices for responsible AI adoption in financial institutions, bringing AI governance, cyber resilience, and financial stability closer together.
The consultation report, published on 10 June, proposes a menu of 12 sound practices that financial institutions could use across organisation-wide AI governance and the AI lifecycle. Responses are open until 22 July 2026.
The report goes beyond model performance or innovation policy. Financial institutions are using AI to transform operations and services, while rapid adoption may also amplify or introduce risks that must be identified and managed. At the financial system level, responsible AI adoption is framed as a way to reduce risks to financial stability.
The proposed practices are aimed at board and senior management decision-making as firms consider business strategy, technology adoption, and risk management in an increasingly AI-enabled environment. The work builds on input from national and regional authorities, standard-setting bodies, financial institutions, and technology vendors.
The consultation asks whether the practices strike the right balance between risks relating to all forms of AI and newer, more complex forms, including generative AI and agentic AI. Agentic systems can act, query, decide, or trigger workflows with varying levels of human supervision. When embedded inside financial services, those actions can touch customer data, payment operations, trading processes, compliance workflows, and supplier environments.
The FSB’s intervention follows a period in which financial regulators have increasingly linked AI adoption to operational resilience. The Bank of England, FCA, European supervisors, and global standard setters have all examined how AI could affect model risk, third-party concentration, fraud, cyber exposure, and market behaviour.
Financial institutions already have mature governance structures for outsourcing, operational resilience, conduct, financial crime, and model risk. AI cuts across those structures. A customer service agent may create conduct risk. A trading assistant may create market risk. A coding agent may introduce software risk. A fraud-detection model may create explainability and fairness concerns. An internal AI tool connected to sensitive data may create confidentiality, identity, and access exposure.
Agentic AI changes the control environment because the risk is not only what the model says, but what the system can do. An AI agent with permissions to access files, open tickets, write code, change configurations, query customer data, or interact with third-party systems needs governance closer to privileged access management than ordinary software approval.
Procurement controls will also become more demanding. Many financial institutions will use AI capabilities through cloud providers, software platforms, analytics vendors, and managed services. Firms will need to know where AI functionality is embedded, which vendors process sensitive data, how models are monitored, and what audit evidence is available.
The UK and Europe will not be insulated from the FSB’s work. The board includes central banks, finance ministries, and regulators from major jurisdictions, and its publications often shape supervisory expectations even where they are not directly binding. UK and European financial institutions should expect AI governance to be assessed alongside operational resilience, outsourcing, technology concentration, and incident response.
The consultation marks another step in AI governance becoming part of mainstream financial resilience. Boards will be expected to understand where AI is used, what it can access, how it fails, which suppliers support it, and whether human oversight exists at points where automated action could create harm.
